Spectral Subtraction for Speech Enhancement

This MATLAB source code implements spectral subtraction for speech enhancement. In the field of speech enhancement, spectral subtraction is a widely-used algorithm that improves speech quality by reducing noise energy through spectral domain processing. The core principle involves subtracting the estimated noise spectrum from the speech signal spectrum to mitigate noise interference. The MATLAB implementation provides an accessible framework for spectral subtraction, featuring customizable parameters for different application scenarios. Key algorithmic aspects implemented include noise estimation during non-speech segments, magnitude spectrum subtraction with over-subtraction factor, and spectral flooring to prevent negative power values. The code structure typically consists of signal framing, FFT transformation, noise spectrum estimation, spectral subtraction operation, and inverse FFT reconstruction. This implementation enables researchers to obtain cleaner, more accurate speech signals, thereby enhancing performance in speech recognition and speech synthesis systems. The modular design allows easy modification of critical parameters like subtraction factor and spectral floor coefficient for optimal results in various noise conditions.